[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82086-en":3,"doc-seo-82086-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},82086,1374391975076,"Riley","https://ap-avatar.wpscdn.com/avatar/14000253ca4ec9f6853?x-image-process=image/resize,m_fixed,w_180,h_180&k=1783305029341752051",8,"Research & Report","HAT Super Resolution and PARSeq+CLIP4STR Voting Ensemble for Extreme In the Wild License Plate Recognition","Entry for the ICIP 2026 Grand Challenge on Extreme In-the-Wild License Plate Super-Resolution (XLPSR) achieving 9.73 wECR on the public validation leaderboard. The approach combines a Hybrid Attention Transformer (HAT) 4× super-resolution front-end with an ensemble of two scene-text recognizers, PARSeq-S and CLIP4STR-B. A confidence-weighted per-character soft-voting scheme performs abstention on uncertain positions using τ=0.33, aligning decisions with the +2/−1/0 scoring rule. Runtime targets 1.7 s per sequence on RTX 3090 within Docker limits.","HAT Super-Resolution and a PARSeq+CLIP4STR  \nVoting Ensemble  \nfor Extreme In-the-Wild License Plate Recognition  \nKarthik Sivarama Krishnan∗†  \n∗ Independent Researcher †Ancestry  \n[ks7585@g.rit.edu](ks7585@g.rit.edu)  \nKoushik Sivarama Krishnan∗  \n∗ Independent Researcher [koushik.nov01@gmail.com](koushik.nov01@gmail.com)  \narXiv :2607 .08896v 1 [ cs .CV] 9 Jul 2026  \n\n| 10 input frames per sequence |  |\n| --- | --- |\n|  |  |\n| Laplacian variance → sharpest reference |  |\n|  |  |\n| ECC affine alignment + pixel-average fusion |  |\n\n\n| HAT ×4 super-resolution (6 views) |  |\n| --- | --- |\n|  |  |\n\nPARSeq-S v2 CLIP4STR-B  \n (32 × 128)   (224 ×224)   \n\n| Per-position [soft-conf. vote](soft-conf. vote) (2:1) + abstention τ=0.33 |\n| --- |\n|  |\n| 7-character output (e.g. AB123CD) |\n\nFig. 1. Pipeline. Ten sequence frames are ranked by Laplacian variance; five aligned crops plus a temporally-fused crop are 4× super-resolved by HAT and passed through two scene-text recognisers whose per-position soft-max outputs are combined by a confidence-weighted vote with τ=0 .33 abstention.  \nAbstract—We describe our entry to the ICIP 2026 Grand Challenge on Extreme In-the-Wild License Plate Super-Resolution (XLPSR), which scored 9.73 wECR on the public validation leaderboard. The system pairs a Hybrid Attention Transformer super-resolution (HAT) front-end with an ensemble of two scenetext recognisers (PARSeq-S and CLIP4STR-B) and a confidenceweighted character-voting scheme that abstains on uncertain positions. We treat XLPSR as a recognition task gated by image legibility: the SR step exists to lift characters out of sub-pixel territory, and the asymmetric scoring rule (+2/−1/0) is exploited explicitly through abstention. Our pipeline runs in 1.7 s per sequence on RTX 3090 (max 2.7 s, p99 2.4 s), well under the 60 s/sequence Docker budget.  \nIndex Terms—License plate recognition, Super-resolution, Scene text recognition, Ensemble, Abstention.  \nI. INTRODUCTION  \nThe XLPSR challenge provides 10 video frames per sequence of moving vehicles with French license plates under extreme in-the-wild degradation: plates as narrow as 12 px, motion blur, JPEG compression, adverse lighting, and occlusion. The task is to predict the plate text. The +2/−1/0 percharacter scoring makes guessing in low-confidence regions negative-EV below P (correct) = 1/3, which we exploit explicitly. Figure 1 summarises the pipeline; the following sections describe each stage.  \nII. METHOD  \nA. Multi-frame fusion  \nEach sample in the blind-evaluation set is supplied as a temporal sequence of already-cropped license-plate images; the public-validation set provides full frames with per-frame bounding boxes which our pipeline crops first. After that branching step the rest of the pipeline is identical: we score every crop by Laplacian variance, pick the sharpest as a reference, align the remaining crops to it with the Enhanced Correlation Coefficient (ECC) algorithm under an affine motion model, and average them pixel-wise into a fused crop. ECC + affine was chosen over optical-flow alternatives because it is robust to the low-contrast conditions that dominate XLPSR crops, requires no ground-truth flow supervision, and is fast enough for CPU-side pre-processing. We carry forward six views per sequence into the recogniser stack: the sharpest reference, up to four additional aligned crops (in input frame order), and the fused crop.  \nB. Super-resolution  \nEach view is upscaled 4 × by Real-HAT-GAN-SRx4 [1], a Hybrid Attention Transformer with six RHAG blocks at embed dim 180, window 16, overlap ratio 0.5, and a PixelShuffle upsampler. We use the public weights without further fine-tuning. Lighter real-time variants such as our earlier SwiftSRGAN [6] would suit tighter latency budgets, but XLPSR’s 60 s/sequence budget lets us prioritise legibility over throughput. SR is the single largest contributor to wECR in our ablation (Table I,+2 .00 over the PARSeq+CLIP4STR no-SR baseline) .  \nC","cbCaisTR7nCv4Sgi","https://ap.wps.com/l/cbCaisTR7nCv4Sgi","pdf",184458,1,2,"English","en",105,"# Introduction\n# Method\n## Multi-frame fusion\n## Super-resolution\n## Two-model OCR ensemble\n## Character voting with abstention\n# Training Data","[{\"question\":\"What is the core pipeline used for XLPSR license plate recognition?\",\"answer\":\"The system ranks video frames by Laplacian variance, aligns multiple crops to a sharp reference using ECC affine alignment, fuses them, then performs 4× super-resolution with HAT. Both the super-resolved views and the original-resolution reference are decoded by PARSeq-S and CLIP4STR-B, and results are merged via confidence-weighted character voting with abstention.\"},{\"question\":\"Why use multi-frame fusion with ECC affine alignment instead of optical flow?\",\"answer\":\"ECC affine alignment is selected for robustness under low-contrast conditions typical of XLPSR crops, avoids requiring ground-truth flow supervision, and remains fast enough for CPU-side pre-processing.\"},{\"question\":\"How does character voting work, and how is abstention applied?\",\"answer\":\"For each character position, soft-max confidence is summed per class across retained recognition tuples, weighting PARSeq twice and CLIP4STR once. The system abstains (emits an underscore) when the winning class soft-max mass falls below τ=0.33, which corresponds to the break-even P(correct)=1/3 implied by the +2/−1/0 scoring rule.\"}]",1784178149,5,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"hat-super-resolution-and-parseqclip4str-voting-ensemble-for-extreme-in-the-wild-license-plate-recognition","",{"@graph":35,"@context":84},[36,52,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,46,49],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":21},"https://docshare.wps.com/document/","Document",{"item":47,"name":12,"@type":42,"position":48},"https://docshare.wps.com/document/research-report/",3,{"item":50,"name":13,"@type":42,"position":51},"https://docshare.wps.com/document/hat-super-resolution-and-parseqclip4str-voting-ensemble-for-extreme-in-the-wild-license-plate-recognition/82086/",4,{"url":50,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":23,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":40,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What is the core pipeline used for XLPSR license plate recognition?","Question",{"text":74,"@type":75},"The system ranks video frames by Laplacian variance, aligns multiple crops to a sharp reference using ECC affine alignment, fuses them, then performs 4× super-resolution with HAT. Both the super-resolved views and the original-resolution reference are decoded by PARSeq-S and CLIP4STR-B, and results are merged via confidence-weighted character voting with abstention.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"Why use multi-frame fusion with ECC affine alignment instead of optical flow?",{"text":79,"@type":75},"ECC affine alignment is selected for robustness under low-contrast conditions typical of XLPSR crops, avoids requiring ground-truth flow supervision, and remains fast enough for CPU-side pre-processing.",{"name":81,"@type":72,"acceptedAnswer":82},"How does character voting work, and how is abstention applied?",{"text":83,"@type":75},"For each character position, soft-max confidence is summed per class across retained recognition tuples, weighting PARSeq twice and CLIP4STR once. The system abstains (emits an underscore) when the winning class soft-max mass falls below τ=0.33, which corresponds to the break-even P(correct)=1/3 implied by the +2/−1/0 scoring rule.","https://schema.org",{"og:url":50,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":50},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,108,113,118,121,126,129,133],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":51,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":28,"doc_module":4,"doc_module_name":45,"category_name":105,"show_sort_weight":106,"slug":107},"Comic",60,"comic",{"id":109,"doc_module":4,"doc_module_name":45,"category_name":110,"show_sort_weight":111,"slug":112},6,"Technology",50,"technology",{"id":114,"doc_module":4,"doc_module_name":45,"category_name":115,"show_sort_weight":116,"slug":117},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":119,"slug":120},30,"research-report",{"id":122,"doc_module":4,"doc_module_name":45,"category_name":123,"show_sort_weight":124,"slug":125},9,"Religion & Spirituality",20,"religion-spirituality",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":124,"slug":128},"World Cup","world-cup",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":130,"slug":132},10,"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":28,"slug":136},19,"General","general"]